Shared Library

Shared Library

The Project/shared/ folder is a lightweight internal library for code that is used by more than one task or the inference app. It prevents duplication without creating tight coupling between tasks. Shared library is kept empty in the template for users to put things in them which are shared for their specific project.

---

The Core Rule

Code lives in shared/ only when it has two or more active consumers — or when you are promoting it there from a task because a second consumer has just appeared.

If you write a utility in Project/TaskA/src/utils/ and later find TaskB needs the same thing, that's the right moment to promote it. Premature sharing creates invisible dependencies that make tasks harder to understand, test, and delete independently.

---

What Belongs in shared/

Good candidates:

Category	Examples
DICOM utilities	Series selection, tag extraction, anonymisation helpers
Common metrics	AUC, Dice, sensitivity/specificity wrappers consistent across tasks
Evaluation tools	Calibration curves, confidence interval bootstrapping
Logging helpers	Standardised MLflow logging wrappers used by all train.py scripts
XNAT utilities	Project/subject/session querying shared across data pipelines

Poor candidates (keep in the task):

Category	Why
Model architectures	Task-specific; premature abstraction leads to over-engineered base classes
Dataset-specific transforms	Tightly coupled to one modality or task
Task config parsing	Each task owns its own config schema

---

Testing Shared Code

Shared code must have its own tests — it is used in multiple places, so a silent bug here affects every consumer. Tests for shared/ live in Project/tests/ (not inside a task's test folder):

Project/tests/
├── __init__.py
├── train_tests/       # Task-specific training tests
└── shared_tests/      # Tests for shared/ utilities

Run with:

cd Project
python -m pytest tests/shared_tests/ -v

---

Last updated: April 2026 · Back to Home